Image Recognition Method for Gas and Liquid Emissions from Electrochemical Energy Storage Cabins Based on an Improved SSD Algorithm

This study addresses the issue of inadequate safety-monitoring methods for lithium-ion battery energy storage systems. An image recognition approach based on a single-shot multibox detector (SSD) algorithm is proposed for detecting gas-liquid emissions within electrochemical energy storage compartments. An experimental platform is developed to simulate the actual operating conditions of lithium-ion battery storage units, and a dataset is constructed from the image data capturing gas-liquid emissions during the overcharging stage. To overcome the limitation of the original SSD algorithm, which features an excessively large model scale that restricts real-time detection, several modifications were implemented: the Visual Geometry Group (VGG) backbone is replaced with MobileNet-V3 to enhance computational efficiency; the squeeze-and-excitation (SE) attention module is substituted with the Coordinate Attention (CA) module to enhance feature extraction capabilities; and mean clustering optimization is applied to refine the default box scale sizes based on the dataset. The experimental results show that the improved SSD model achieves a 92.2% reduction in model size (from 91.9 MB to 7.2 MB), with a 1.54% increase in average accuracy (from 90.38% to 91.92%). The prediction speed increased from 15 to 58 frames per second (FPS), meeting the real-time detection requirements for lithium-ion battery energy storage compartments.